How an AI startup founded by former DOGE staffers broke into Anthropic

Researchers at cybersecurity startup Cathedral exploited a vulnerability in third-party enterprise software used by Anthropic to enter its internal sales, HR and IT systems and access sensitive data; Anthropic said it cut off access the same day. According to Cathedral's report, researchers running AI agents found a path in under 12 hours and accessed meeting records, salary and hiring information, and Slack messages discussing its models. Anthropic said only a small subset of systems was accessed, and the researchers were unable to reach messages users sent to models, model responses, or model code or weights; company monitoring had already detected some activity before the report, and access was shut down the same day. Anthropic said there is no evidence anyone other than Cathedral exploited these vulnerabilities, has disabled the relevant enterprise software, and re-enabled it after the vendor confirmed a fix. Cathedral disclosed the findings through Anthropic's bug bounty program, and Anthropic has notified federal officials.